LiZAD: A Lightweight Zero-Shot Anomaly Detection Framework for Industrial Manufacturing

Technical Summary: LiZAD – A Lightweight Zero-Shot Anomaly Detection Framework

Problem Statement
In modern high-throughput industrial manufacturing, product configurations and visual characteristics change frequently. Traditional supervised anomaly detection requires extensive data collection and annotation for every new scenario, which is often costly and infeasible. While Zero-Shot Anomaly Detection (ZSAD) offers a solution by detecting defects in unseen categories without target-specific training data, existing ZSAD approaches face a critical bottleneck: they are computationally intensive. Built on large vision-language backbones, these models demand high memory and processing power, rendering them unsuitable for deployment on resource-constrained edge devices or real-time production lines where inference speed is paramount.

Methodology
The authors propose LiZAD, a lightweight framework designed specifically for real-time ZSAD on edge devices. The architecture addresses the trade-off between computational efficiency and fine-grained localization through three core components:

1. Visual Backbone (DINOv3): Instead of standard CLIP-style visual encoders which are optimized for global semantic alignment, LiZAD utilizes DINOv3 (specifically the compact ViT-S/16 variant). This choice is driven by DINOv3's ability to provide dense, spatially aware visual features crucial for precise pixel-level anomaly localization. The model uses a frozen encoder to extract global tokens and patch-level embeddings.
2. Text Encoder (MobileCLIP2): For the text branch, the framework employs MobileCLIP2, an architecture engineered for low-latency execution on edge hardware. It utilizes a highly optimized convolution-transformer design and multi-modal reinforced training to extract expressive text embeddings with minimal memory footprint.
3. Shared Latent Space & Projection: Since the visual and text encoders operate in different dimensional spaces and are not jointly trained, LiZAD introduces lightweight, trainable linear projection heads. These heads map the visual patch embeddings, global tokens, and text embeddings into a shared latent space.
   • Prompting: The system uses standard templates ("an image of a {state} {object}") to generate normal and abnormal text embeddings.
   • Scoring: Anomaly scores are computed via cosine similarity between projected visual patches and the normal/abnormal text embeddings. A softmax function converts these similarities into anomaly probabilities.
   • Regularization: A global-context regularizer based on the projected global token is added to stabilize the embedding space and calibrate directions beyond sparse patch-level supervision.
   • Output: The resulting coarse anomaly map is upsampled via bilinear interpolation to generate a dense, pixel-level anomaly heatmap.

Key Contributions

• First Lightweight ZSAD for Edge: LiZAD is presented as the first fast, lightweight, and accurate ZSAD framework explicitly built for real-world deployment on resource-constrained devices.
• Architectural Efficiency: By pairing DINOv3 with MobileCLIP2 and using only lightweight projection heads for training (keeping encoders frozen), the framework drastically reduces computational overhead.
• Real-World Validation: The framework was successfully deployed on NVIDIA Jetson NX and Jetson AGX edge devices and tested on a real production line at the Industrial Computer Engineering Laboratory (ICE Lab) at the University of Verona.

Experimental Results
The authors evaluated LiZAD against six state-of-the-art ZSAD models (CLIP, WinCLIP, APRILGAN, AnomalyCLIP, AdaCLIP, Bayes-PFL) across four industrial datasets: VisA, BTAD, MPDD, and MVTec-AD.

• Efficiency Gains:
  • Memory: Reduced by an average of 61.5% (down to 1.01 GB).
  • Parameters: Reduced by 74.6% (down to 97.44 M parameters).
  • Latency: Achieved a 3.02× speedup (69.9% reduction in inference latency) compared to the average of competing models. On Jetson NX, it achieved 754.6 ms per sample; on Jetson AGX, 554.4 ms.
• Performance Trade-off: Despite the significant reductions in complexity, LiZAD maintains competitive detection performance.
  • It achieved an average pixel-wise AUROC of 88.05%.
  • This represents a drop of only 6.4% relative to the best-performing state-of-the-art model across the tested datasets.
  • Specifically, it remained within 1.3% of the best result on VisA and 5.2% on MVTec-AD.
• Qualitative Results: Visual comparisons demonstrate that LiZAD produces sharp and concentrated anomaly responses across diverse textures and defect patterns, confirming its robust localization capabilities.

Significance
The paper claims that LiZAD successfully bridges the gap between high-accuracy zero-shot anomaly detection and the strict real-time, resource-constrained requirements of industrial environments. By demonstrating that substantial reductions in memory and parameters can be achieved with only a minimal trade-off in localization accuracy, the framework enables the deployment of advanced anomaly detection systems on edge platforms where previous heavy models were impractical. The successful testing on a real production line underscores its potential for immediate industrial application.